生成对抗网络在油品分子组成预测方面的研究  

Research on Molecular Composition Prediction of Oil Products Based on Generative Adversarial Network

作　　者：宋亚峰 王杭州[2] 谷明章 王弘历[2] 商辉[4] 孙宝文[2] 纪晔[2] SONG Yafeng;WANG Hangzhou;GU Mingzhang;WANG Hongli;SHANG Hui;SUN Baowen;JI Ye(Sinopec Tianjin Petrochemical Company;PetroChina Planning and Engineering Institute;PetroChina Liaohe Petrochemical Company;China University of Petroleum(Beijing))

机构地区：[1]中国石油化工股份有限公司天津分公司 [2]中国石油天然气股份有限公司规划总院 [3]中国石油天然气股份有限公司辽河石化分公司 [4]中国石油大学(北京)

出　　处：《油气与新能源》2021年第4期90-100,共11页Petroleum and new energy

摘　　要：炼油行业呈现由粗放向精细发展趋势。分子炼油技术已成为当前研究的热门方向,而实现分子炼油技术的第一步就是获取分子组成。在定义了含590种常见汽油组分的分子库为预测模型输出端的基础上,以18种汽油油品宏观物性为输入端,以某炼厂31组汽油检测报告为基础数据,随机取其中29组数据为模型训练集数据库,剩余两组数据作检测组生成预测结果,最终开发了以生成对抗网络为原理预测油品分子组成的模型。从原数据库中抽取两组数据为新的检测组,再将原检测组两组数据归为训练集数据库,分三次验证模型效果,并选取油品的主要分子组成将预测值与实际值进行对比。结果显示,第一组生成对抗网络模型预测结果的平均误差为5.80%,3.86%;第二组为4.85%,3.11%;第三组为3.86%,3.07%。生成对抗网络法模型所得预测值与实际值基本相同,由此认为模型的模拟结果较好地反映了汽油的分子组成,为今后获取重油的分子组成提供了一种方法。The refining industry presents a trend from extensive to fine in its development.Molecular refining technology has become a hot research topic,and the first step to realize molecular refining technology is to obtain molecular composition.Defining a molecular library containing 590 common gasoline components as the output of the prediction model,the research used 18 kinds of gasoline oil macroscopic properties as the input.31 groups of gasoline test reports of a refinery were taken as the basic data,29 groups of data were randomly selected as the model training set database,and the remaining two groups of data were used as the test group to generate prediction results.Finally,a model was developed to predict the molecular composition of oil products based on the principle of generative adversarial network.Two groups of data were extracted from the original database as the new detection group,and then the two groups of data of the original detection group were classified into the training set database.The model effect was verified three times by comparing predicted results to actual values based on main molecular compositions.The findings showed that the average error of the prediction results of the generative adversarial network model was 5.80%and 3.86%for the first group,4.85%and 3.11%for the second group,and 3.86%and 3.07%for the third group.The predicted value obtained by the generative adversarial network method is basically the same as the actual value.Therefore,it is considered that the model simulation results can well reflect the molecular composition of gasoline,which also provides a method for obtaining the molecular composition of heavy oil in the future.

关 键 词：分子管理 油品分子组成预测模型 生成对抗网络法 

分 类 号：TE62[石油与天然气工程—油气加工工程]